AI agent runtime authorization

Agents are already inside your data. TrustAI decides what they can do.

Deploy more AI agents without increasing data exposure: TrustAI governs what each agent can do, which MCP tools it can use and what data it can see, based on who it's acting for and why.

✕
With a kill switch for every agent.
Talk to an architect
Talk to an architect
★★★★★4.8Gartner Peer Insights
Decide once. Enforce where your systems already are.
01
Agent
Initiates
02
Agent
Delegates
03
MCP tools
Act on data
Identity security
User · Agent · Delegation
Okta
Microsoft Entra ID
Intent
Approved purpose · Out of scope
Data security
Sensitivity · Entitlements
TrustAI
Runtime authorization
Allow
Allow with limits
Deny
One decision on every governed call
✕ Kill switch per agent
Enforced through the systems you already run
Agent platform plug-ins
In the agent runtime
Microsoft Foundry
AWS Bedrock
Copilot Studio
Snowflake Cortex
Databricks
AI and MCP gateway integrations
Your existing gateways
Kong
AWS AgentCore Gateway
Gravitee
Native data controls
At the source
Enterprise data, wherever the workflow goes
Snowflake
Databricks
SQL Server
Power BI
Amazon S3
PostgreSQL
Dremio
SharePoint (MCP tools)
Azure · AWS · On-premises
Illustrative architecture. Enforcement paths vary by integration.
Trusted by data teams at
Best Egg
Partners
Why TrustLogix

Permission to connect.
Not permission to do everything.

Your teams are already putting agents into production, with access set up simply to connect. TrustLogix closes that gap from the data up.

01

Starts from the data

Policy built on what's sensitive and who can reach it.

02

Not another gateway

Decisions enforced in the systems you already run.

03

One decision

Identity, data and purpose weighed on every call.

One request. A precise decision.

Let the task succeed.
Keep the data within bounds.

Change who's asking or what they ask. Same agent, different answer.

The request
FA
Finance copilot
Same agent for both analysts
Who's asking
EMEA finance analyst
US finance analyst
Switch to the US finance analyst: same agent, same question, US rows instead.
Approved purpose
Regional revenue reporting
Choose a request
Revenue and customer details
Includes contact and payment information
Regional revenue totals
Stays within the approved purpose
Export payment records
Outside the approved purpose: whole call denied
"Summarise Q3 revenue by customer, including contact and payment details."
TrustAI runtime decision
✓ Allow with limits

The summary goes through.
The excess access doesn't.

Why it's asking
Purpose
Within regional revenue reporting ✓
Revenue by customer
Person
EMEA records only ✓
Contact email
Data label
Personal information masked ◆
Payment card details
Data label
Access denied ✕
Governed response · Synthetic data
Customer
Region
Contact
Payment card
Customer A
EMEA
••••@••••
Withheld
Customer B
EMEA
••••@••••
Withheld
Why this decision?
Agent
Finance copilot (Azure AI Foundry)
Acting for
EMEA finance analyst (Entra ID group Finance EMEA)
Purpose
Regional revenue reporting. Request matches the approved intent
Policy
Finance agents: own region only, mask personal data, no card data
Data labels
PII from Purview, PCI from Snowflake tags
Enforced in
Snowflake, as a native row access and masking policy
Four checks. One policy control plane.

Govern the agent, the tool, the data
and the purpose.

So every agent can do its job, and only its job.

01 · Agent

Who can act?

Who the agent acts for, what it's approved to do, and the limits of that authority.

Agent identity · Delegation · On behalf of
02 · Tool

What can it use?

Access to MCP servers and to individual tools, per agent, so a permitted connection doesn't become unrestricted use.

MCP servers · Tools · Actions
03 · Data

What can it see?

Policy on records, fields and documents, including row filtering and column masking where supported.

Rows · Columns · Documents
04 · Purpose

Why is it asking?

Each call is checked against what the agent is for. Two agents with the same access get different answers when their purposes differ.

Intent · Approved purpose · Out of scope

Reuse business rules across connected systems.
Not a new authorization model for every agent.

Explore TrustAI
Explore TrustAI
Intent-based access control (IBAC)

Know what each agent is for.Hold every call to it.

Identity tells you who the agent is. It doesn't tell you what the agent is for. TrustAI works that out first, then checks every call against it.

01

Understand the agent

TrustAI reads the agent's instructions and tools and drafts its intent: what it's for, the data it needs and what's out of scope.

02

Recommend the policy

It turns that intent into tool level allow and block recommendations, each with a reason. Your team reviews and approves them.

03

Decide every call

At runtime, each tool call is checked against the approved intent, the person the agent acts for and the labels on the data.

FC
Finance copilot
Agent intent · Azure AI Foundry
✓ Approved
1
Intent, drafted from the agent
Purpose
Regional revenue reporting
In scope
Revenue totals and revenue by customer, for the analyst's region
Out of scope
Payment card data · Bulk exports
2
Recommended tool policy
query_revenue
Needed for regional reporting
Allow
lookup_customer
Needed for revenue by customer
Allow
export_payment_records
Outside the approved purpose
Block
3
At runtime
"Export all Q3 payment records to a spreadsheet"
✕ Denied: outside purpose
From coverage to containment

Know what is governed.
Act when it goes off course.

See connected agents and their data access, find coverage gaps, and apply least privilege policies.

01

Establish coverage

Tell discovered agents apart from access that is governed.

02

Evaluate the request

Check approved purpose and policy, with a decision trail.

03

Contain selectively

A kill switch per agent revokes its governed access without shutting down the agents that are behaving.

04

Monitor what agents touched

The Guardian agent watches what agents actually did with your data, and tags findings for compliance.

Access control
Illustrative view
F
Finance copilot
Regional revenue reporting
● Active
Governed scope
EMEA revenue · Read only
R
Research agent
Approved document analysis
● Active
Governed scope
Approved documents · Read only
!
A request needs containment. The research agent keeps asking for payment records. That's outside its approved purpose.
Kill switch: revoke this agent
Governed access only. A person presses the kill switch.
Guardian agent

See what agents actually touched

Guardian reads the platforms' own logs, not a copy of the traffic, and flags what needs a closer look.

Sensitive data access
Access outside an agent's purpose
Unused permissions
Data movement
Compliance tagging

Findings mapped to the frameworks you report on

Each Guardian finding is tagged to the framework it relates to, so governance teams can report on agents without rebuilding the evidence.

EU AI Act
ISO 42001
Version -1
Works with your architecture

Keep your stack.
Make every part of it count for more.

Agents don't live in one platform. A single task can cross your identity provider, an AI gateway and three data platforms.

Each of those tools has controls of its own, and none of them sees the whole call. TrustLogix plugs into them and connects what they already know into one decision, without adding a new hop to the traffic path.

Identity provider

Entra ID
Okta
On its own
Knows who the person and the agent are.
With TrustLogix
Those identities and groups decide what each agent can do with data. No second directory to run.

Classification and catalogs

Purview
Unity Catalog
Snowflake tags
Collibra
Alation
On its own
Labels what's sensitive.
With TrustLogix
Those labels are enforced for people and agents, wherever the data is queried.

AI security tools and gateways

AI gateways
MCP gateways
Prompt security
On its own
Inspects prompts and responses.
With TrustLogix
Keeps inspecting. We add the decision on whether the call should happen, for this person and this purpose.

Data platforms

Snowflake
Databricks
SQL Server
Power BI
On its own
Holds the data and its access controls.
With TrustLogix
Policy runs as the platform's own controls, so you keep its performance and its audit trail.
Pick a tool to see what changes for it.
Your tools + TrustLogix
One decision across all of them. Nothing to rip out or replace.
Identity & delegation
Approved purpose
Data labels
One runtime decision. Enforced where your systems already are.
01

Agent platform plug-ins

Plug into the platforms where agents are built and run, and apply policy per agent, per tool and per person. No central proxy.

02

AI and MCP gateway integrations

Add the authorization decision to the LLM and MCP gateways you already run.

03

Native data controls

Enforce fine-grained access through the data platform's own controls.

04

Your applications

TrustApp: your apps call one REST API for a decision before they reach data, an API or an LLM.

Explore your platform
Microsoft SnowflakeDatabricksAWS
Start with one agent workflow. Extend coverage as your AI program grows.
Version -2
Works with your architecture

Keep your stack.
Make every part of it count for more.

Agents don't live in one platform. A single task can cross your identity provider, an AI gateway and three data platforms.

Each of those tools has controls of its own, and none of them sees the whole call. TrustLogix plugs into them and connects what they already know into one decision, without adding a new hop to the traffic path.

On their own

Four tools. Four separate checks.

Each one is right about its own piece. None of them sees the whole call.

Identity provider
Knows who the person and the agent are.
Decides alone
Classification and catalogs
Labels what's sensitive.
Decides alone
AI security tools and gateways
Inspects prompts and responses.
Decides alone
Data platforms
Holds the data and its access controls.
Decides alone
With TrustLogix

One decision across all of them.

Same tools, now connected. Nothing to rip out or replace.

Identity provider
Those identities and groups decide what each agent can do with data. No second directory to run.
Entra ID · Okta
Classification and catalogs
Those labels are enforced for people and agents, wherever the data is queried.
Purview · Unity Catalog · Collibra
AI security tools and gateways
Keeps inspecting. We add the decision on whether the call should happen, for this person and this purpose.
AI gateways · MCP gateways
Data platforms
Policy runs as the platform's own controls, so you keep its performance and its audit trail.
Snowflake · Databricks · SQL Server
TrustLogix
One
decision
Trusted context
Identity & delegation
Approved purpose
Data labels
01

Agent platform plug-ins

Plug into the platforms where agents are built and run, and apply policy per agent, per tool and per person. No central proxy.

02

AI and MCP gateway integrations

Add the authorization decision to the LLM and MCP gateways you already run.

03

Native data controls

Enforce fine-grained access through the data platform's own controls.

04

Your applications

TrustApp: your apps call one REST API for a decision before they reach data, an API or an LLM.

Explore your platform
Microsoft SnowflakeDatabricksAWS
Start with one agent workflow. Extend coverage as your AI program grows.
What changes for the business

Say yes to more agents.
Without saying yes to more exposure.

Written for the people who fund the AI program, not just the people who secure it.

Speed

Get AI projects through security review sooner

Each agent gets the access its task needs and nothing more, so security can approve the next agent instead of holding it back.

How Runtime authorization per agent, per tool and per person
Risk

Stop sensitive data exposure as it happens

Every call is checked in real time. Data an overreaching agent shouldn't see is masked or denied before it's returned, not found in a log afterwards. Revoke one agent and the rest keep working.

How Runtime policy enforcement and a kill switch per agent
Effort

Get more from the tools you already own

Reuse your identity provider, data labels and existing data policies. One policy for people and agents, with nothing to rip out or replace.

How Policy consolidation across Snowflake, Databricks and Microsoft
Audit

Have the evidence ready when audit asks

Every decision is recorded with the agent, the person, the purpose and the policy, so you can show why access was granted, limited or refused.

How A decision record on every governed call
50% faster data access
and 25% faster audits at Jefferies
90% fewer access misconfigurations
at Ericsson
What customers have seen
Data access results on the TrustLogix platform, the same policy foundation TrustAI extends to agents.
A clear path to your first workflow

Your stack.
Your questions.

Explore the platform, or work through the architecture with our team.

Do we have to replace any of our tools?

No. TrustLogix plugs into your identity provider, classification and catalogs, AI security tools and data platforms, and connects what they already know into one decision. See how each one does more with TrustLogix.

Is TrustAI another gateway?

No. TrustAI isn't another gateway in the traffic path. A gateway inspects prompts and responses as traffic passes through. TrustAI decides whether a call should happen at all, for this person and this purpose, and that decision is enforced through agent platform plug-ins, your AI and MCP gateways, or natively in the data platform.

How does TrustAI know what an agent is for?

With intent-based access control. TrustAI reads the agent's instructions and tools, drafts its intent, and turns that into tool level recommendations. Your team reviews and approves them before they're enforced.

Does our data pass through TrustLogix?

With native data controls, policy runs inside the data platform, so query results stay there.

Is the decision always allow or deny?

No. A request can be allowed with limits, such as a regional row filter or masked personal information, where the enforcement integration supports it. The task can succeed without exposing all the data the connection can reach.

What does the kill switch actually revoke?

It revokes an agent's governed access on supported enforcement paths. It does not terminate the agent process, reverse completed actions, or control credentials and paths outside TrustAI coverage.

Where should we start?

Bring one agent, one sensitive dataset and one real task. Together we map the permissions the task needs, identify data that should stay out of reach, and confirm the right enforcement path.

One agent.
One sensitive dataset.
One real task.
Pick one agent and one sensitive dataset. We set it up with you on a pre-production copy, using the identity, labels and platforms you already have, and show which permissions the task needs and what stays out of reach.